Work IQ GA Turns Microsoft 365 Context Into a Metered Agent Primitive

Microsoft just put a meter on one of the most valuable ingredients in enterprise AI: the context of work itself.

Work IQ, Microsoft’s API layer for grounding agents in Microsoft 365 context, reaches general availability on June 16 with consumption-based pricing through Copilot Credits, according to Microsoft Partner Center. That sounds like a billing-note footnote until you look at what Work IQ actually exposes: mail, calendar, files, people, chats, sites, agent-to-agent interfaces, a redesigned remote MCP server, REST APIs, and SharePoint Embedded workspaces for persistent agent state inside the tenant boundary.

The short version: Microsoft 365 context is no longer just something Microsoft 365 Copilot gets to use. It is becoming a platform primitive that custom agents, Foundry apps, Copilot Studio workflows, and third-party AI systems can call. The catch is that once you turn organizational memory into an API, somebody has to pay per call.

That is the right direction. It is also where enterprise AI costs go to become interesting in the accounting sense, which is rarely the fun kind of interesting.

The context layer is the product

Microsoft’s developer post frames Work IQ as four pieces: chat, context, tools, and workspaces. The important phrase is not “chat.” We have plenty of chat. The important part is a managed layer that assembles Microsoft 365 context, lets agents act through a compact tool surface, and keeps intermediate state in SharePoint Embedded rather than some random vector database a prototype picked because it had a good README.

The new GA endpoints include A2A for agent-to-agent collaboration, REST for application integration, and a redesigned remote MCP server. Microsoft says the Work IQ MCP surface collapses hundreds of operations into 10 generic tools covering Microsoft 365 data such as mail, calendar, files, people, chat, and sites. Instead of handing the model a giant catalog of narrowly named functions, Work IQ uses simpler verbs like fetch, create, and update, with resource paths defining what the agent is touching.

That is a serious architectural choice. LLM agents are bad at choosing from enormous tool menus. They are better when the action space is small, composable, and discoverable. Microsoft’s getSchema capability follows the same logic: let the agent discover data shape at runtime instead of requiring every integration to become another bespoke plugin.

This is Microsoft doing the very Microsoft thing that often wins enterprise platforms: take the messy thing everyone is building one-off, give it a control plane, connect it to identity, wrap it in admin controls, and then meter it.

Copilot Credits are now part of the architecture review

Partner Center says there is no separate Work IQ API subscription, SKU, or per-user license. Usage is billed through Copilot Credits. Microsoft 365 Copilot prebuilt agents get Work IQ covered by existing licensing, but custom agents built in Copilot Studio, Foundry, or third-party AI platforms incur consumption when they ground themselves in Microsoft 365 data through Work IQ APIs.

That distinction matters. Microsoft’s own agents stay bundled and predictable. Your custom agent becomes a cloud workload.

The licensing page gives illustrative ranges: light scenarios at $0.20 to $0.40 per call, medium scenarios at $0.30 to $0.75, and heavy scenarios at $0.50 to $1.50. Those prices are not alarming if a person asks an agent for a manager-task checklist once or twice. They become very real when a workflow runs across every account, every seller, every morning, with long email threads, customer docs, calendar context, and follow-up actions.

This is the part engineering teams need to stop treating as procurement trivia. Agent design now includes marginal cost design. How much context does the agent retrieve? How often? Does it cache intermediate results? Does it call Work IQ once per user request, once per subtask, or inside a loop that looked harmless in staging? Does the agent know when a narrow Graph call is enough, or does it default to broad Work IQ grounding because that made the demo look smarter?

“Consumption-based pricing” is not automatically bad. It can be more honest than hiding everything behind a seat license. But it moves the failure mode. The bad outcome is no longer just an agent that gives a mediocre answer. It is an agent that gives a mediocre answer after spending real money retrieving context it did not need.

The governance model is promising, but only if teams use it

Work IQ’s security model is more interesting than the billing model. Microsoft says the redesigned MCP server uses broad high-level permissions plus a Rego-based policy engine for fine-grained enforcement. Requests are user-scoped and evaluated against resource paths, request methods, identity, and data content. Tool invocations are logged and evaluated, with support for auditability, usage analytics, rate limiting, and compliance enforcement.

That is the shape enterprise agents need. The old model — collect OAuth scopes until the demo works, then hope the app review catches the scary parts — does not survive contact with agents that can reason across mail, meetings, Teams chats, files, and people. If an agent can assemble the work graph, the relevant security boundary is not one document or one prompt. It is the chain of retrievals and actions the agent performed on behalf of a user.

Partner Center’s admin guidance is therefore not optional ceremony. Microsoft says admins must enable consumptive billing before Work IQ usage begins, then configure payment methods, access policies, usage limits, and alerts in the Microsoft Admin Center. Translation: if your organization turns this on without per-agent budgets, environment separation, and log review, you did not deploy infrastructure. You opened a tab.

The first production checklist should be boring by design. Inventory every custom agent that reads Microsoft 365 data. Decide whether each needs Work IQ or narrower Graph access. Put preview, internal production, and third-party agents under separate policies. Start with budgets lower than the business thinks it needs. Alert on call spikes, heavy-scenario skew, and unusual department access. Log tool invocations somewhere security and platform teams actually query.

Then run the ugly test, not the demo test: stale meetings, duplicate docs, ambiguous permissions, long email threads, conflicting customer records, and a user asking for a recommendation with receipts. If Work IQ can ground that workflow clearly, stay inside policy, and produce a cost profile the business understands, it earns its keep. If not, the team has learned something before the invoice teaches it more loudly.

Microsoft is turning agent lock-in into agent infrastructure

The strategic move is obvious and effective. Work IQ makes Microsoft 365 context easier for agents to use, but it also makes Microsoft’s work graph the default substrate for enterprise agents. Foundry, Copilot Studio, MCP, A2A, SharePoint Embedded, Microsoft 365 admin controls, and Copilot Credits all pull in the same direction: build your agents where the organizational context already lives.

There is nothing inherently wrong with that. Most enterprises would rather buy a governed context layer from Microsoft than assemble one from search indexes, Graph calls, vector stores, custom permission filters, and vibes. But builders should be honest about the trade. Work IQ reduces integration friction by concentrating context, policy, telemetry, and billing in one place. That concentration is exactly what makes it useful — and exactly what makes it sticky.

The practical editorial take: Work IQ GA is not “another Microsoft 365 API.” It is Microsoft turning work context into a metered agent substrate. Use it when the agent genuinely needs rich organizational grounding. Avoid it when narrower data access will do. And before any custom agent gets production access, review it the same way you would review a new cloud service: permissions, observability, budgets, failure modes, and who owns the pager when the very helpful robot starts calling the expensive endpoint too often.

Sources: Microsoft Partner Center, Microsoft 365 Developer Blog, Microsoft Licensing, microsoft/work-iq, GitHub Copilot billing docs